Improved deep learning model for differentiating novel coronavirus pneumonia and influenza pneumonia

Min Zhou; Yong Chen; Dexiang Wang; Yanping Xu; Weiwu Yao; Jingwen Huang; Xiaoyan Jin; Zilai Pan; Jingwen Tan; Lan Wang; Yihan Xia; Longkuan Zou; Xin Xu; Jingqi Wei; Mingxin Guan; Jianxing Feng; Huan Zhang; Jieming Qu

During developing the deep learning model, the first problem we met lies in transferability (5, 6) . The model performs better on cases from CT device appearing in the training set than cases from CT devices not included. To address this problem, especially when classifying image data from multiple CT devices, we proposed a Trinary classification scheme to penalize the network from extracting device specific features during learning. By doing so, it would lead to high cost on the random region inputs, forcing the model to extract more lesion specific features. Although it is impossible to exclude all device specific features, we observed a visible improvement in performance (AUC from 0.85 to 0.89) on patient level classification. Such a performance is comparable with the judgement of experienced specialists. 13 (22.8%) NCP patients presenting uncommon CT findings, such as a small ground-glass opacity (GGO) in the central part were correctly classified by our Trinary scheme, . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.